Learning discriminative prototypes: Adaptive relation-aware refinement and patch-level contextual feature reweighting...

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Few-shot learning (FSL) addresses classification with limited labeled data, crucial for data-scarce scenarios.
• Metric-based FSL relies on class prototypes, but existing methods struggle with static feature selection and lack of fine-grained local feature expression.
• Global features in prototypes limit representational power due to data scarcity and quality issues.

Purpose of the Study:

• To propose a novel framework, Learning Discriminative Prototypes (LDP), to overcome limitations in existing FSL methods.
• To enhance the discriminative power and robustness of class prototypes in few-shot classification.
• To improve model generalization in data-scarce environments.

Main Methods:

• Introduced Learning Discriminative Prototypes (LDP) framework with two key modules.
• Implemented Adaptive relation-aware refinement to dynamically model inter-class prototype relationships and enhance feature robustness.
• Developed Patch-level contextual feature reweighting to obtain more discriminative prototypes through local feature interactions.

Main Results:

• LDP demonstrated strong competitiveness across five diverse datasets (standard and cross-domain).
• Achieved over 12% accuracy improvement in 1-shot settings on miniImageNet and tieredImageNet compared to baseline methods.
• Showcased a 6.45% accuracy improvement on the cross-domain CUB200 dataset in the 1-shot scenario.

Conclusions:

• LDP effectively enhances prototype representation for few-shot learning.
• The proposed framework significantly improves classification performance, especially in data-scarce and cross-domain settings.
• LDP offers a robust solution for addressing model generalization challenges in FSL.